Learned behavior generalizes to different noise levels.

The network was trained on two noise levels (0% and 75%, red dots). During testing all 6 noise levels were shown (A) The agent achieves good performance on noise levels that were not seen during training (black dots). (B,C) The agent also learns to use an adequate number of observations on noise levels not seen during training (black dots (A), dashed lines (C)), indicating it has really learned a mechanism to integrate information and make a decision when enough evidence is collected.